set readme to add images for template requirements

Author: rcghpge

README.md

@@ -0,0 +1,260 @@
+# version-cv
+
+* **One Sentence Summary** 
+
+  This repository symbolic math classification using image-based, text-based, and classical learning models including EfficientNetB0, GPT2, Word2Vec, and XGBoost.
+
+---
+
+## Overview
+
+* **Definition of the tasks / challenge** 
+  The project explores symbolic reasoning and multimodal classification using visual and textual inputs. A binary classification task was constructed using the Hugging Face version of the MathWriting dataset by shuffling LaTeX ID and image pairings to generate synthetic binary labels (0 or 1).
+
+* **Approach** 
+  Four models were developed and compared:
+  - `EfficientNetB0` for visual symbol classification
+  - `GPT2` fine-tuned on LaTeX token sequences
+  - `Word2Vec` for symbolic embeddings
+  - `XGBoost` applied to Word2Vec vectors
+
+* **Summary of the performance achieved** 
+  - EfficientNetB0: ~51% accuracy 
+  - GPT2: ~47% accuracy 
+  - Word2Vec: ~48% macro F1 score 
+  - XGBoost: ~49% macro F1 score 
+  - GPT2 prototype: ~76.1% accuracy, ROC AUC score .832 (83.2%)
+
+---
+
+## Summary of Workdone
+
+### Data
+
+* **Data:**
+  * **Type:**
+    - Image: 224×224 LaTeX-rendered symbols (PNG)
+    - Text: Tokenized LaTeX expressions
+    - Output: Multi-class and binary labels
+  * **Size:**
+    - ~4,000 total samples (multi-class)
+    - ~2,000 binary samples from MathWriting
+  * **Split:**
+    - Multi-class: 2800 train / 600 test / 600 val 
+    - Binary: 1600 train / 200 test / 200 val
+
+#### Preprocessing / Clean up
+
+* Image resizing and normalization 
+* GPT2 tokenization of LaTeX strings 
+* Shuffling LaTeX ID–image pairs to generate 0/1 binary labels 
+* Oversampling used for class balancing
+
+#### Data Visualization
+
+* Token length distributions, symbol histograms 
+* t-SNE clustering of Word2Vec vectors 
+* Confusion matrices and ROC curves for binary task
+
+### Problem Formulation
+
+* **Input / Output:** 
+  - Input: math symbol image or LaTeX sequence 
+  - Output: categorical label or binary label
+
+* **Models:** 
+  - `EfficientNetB0` (image classifier) 
+  - `GPT2` (LaTeX sequence transformer) 
+  - `Word2Vec` + `XGBoost` (vectorized + classical)
+
+* **Loss, Optimizer, Hyperparameters:** 
+  - Loss: Categorical/Binary Crossentropy 
+  - Optimizers: Adam, AdamW 
+  - Batch Size: 32, Epochs: 30, Early stopping applied
+
+---
+
+### Training
+
+* **Software & Hardware:** 
+  - Python 3.12.11 
+  - TensorFlow 2.15, 2.18, and 2.19, HuggingFace Transformers, Gensim, XGBoost 
+  - Trained on: Dell Precision Workstation 5510, Lambda Cloud compute - NVIDIA RTX A6000, NVIDIA Quadro RTX 6000, NVIDIA GH200 Grace Hopper Superchip, NVIDIA A100 Tensor Core GPU, and NVIDIA A10 Tensor Core GPU 
+
+* **Training time:** 
+
+  - Full training, tuning, and experimentation across all models took a couple of weeks 1-2.
+
+* **Training curves:** 
+
+  - Models converged within 10–20 epochs 
+  - Validation monitoring and checkpointing used 
+  - Binary classifier reached high AUC after 5–8 epochs
+
+* **Challenges:** 
+
+  - GPT2 token memory and batch size constraints 
+  - Symbol label imbalance and binary class noise 
+  - Feature drift during Word2Vec–XGBoost transfer
+  - Scoping the work only to what is understood as data science today
+
+
+### Performance Comparison
+
+* **Metrics:** Accuracy, Macro F1, ROC AUC (binary)
+
+| Model                  | Accuracy | F1 Score (macro) |
+|-----------------------|----------|------------------|
+| EfficientNetB0        | 51%      | NaN              |
+| GPT2                  | 47%      | 0.34             |
+| Word2Vec              | 48%      | 0.32             |
+| XGBoost               | 49%      | 0.48             |
+| GPT2 Prototype        | 76.1%    | 0.76             |
+
+---
+
+* **Visualizations:** 
+  ![](assets/modelcomparisons.png) 
+  ![](assets/gpt2confusionmatrix.png) 
+  ![](assets/prototypesroccurve.png)
+
+---
+
+### Conclusions
+
+* EfficientNetB0 was optimal for symbolic image classification 
+* GPT2 achieved strong results for token sequence inputs 
+* MathWriting binary label task showed robust generalization from shuffled samples 
+* Word2Vec + XGBoost was lightweight and interpretable
+* GPT2 prototype outperformed all model builds (was not fully built)
+
+---
+
+### Future Work
+
+* Integrate CROHME-style handwritten data 
+* Explore layout-aware ViT + LSTM hybrids 
+* Build MathOCR pipeline with MathPix-style parsing 
+* Expand symbolic corpora beyond MathWriting
+* Explainable AI (XAI)
+* Secure ML/DL systems
+* Responsible data science
+
+---
+
+## How to reproduce results
+
+```bash
+git clone https://github.com/rcghpge/version-cv.git
+cd version-cv
+pixi install  # Or: pip install -r requirements.txt
+```
+
+## Launch Jupyter Notebooks
+
+```bash
+jupyter notebook notebooks/
+```
+
+Notebook index:
+
+ * `basemodel.ipynb`: basel model + EfficientNetB0 backbone model
+ * `model2.ipynb`: GPT2 model
+ * `model3.ipynb`: Word2Vec training + classical classifier
+ * `model4.ipynb`: XGBoost integration
+ * `mathwriting.ipynb`: prototyping builds
+ * `comparemodels.ipynb`: side-by-side model analysis
+ * `prototypes.ipynb`: GPT2 prototype (performed the best)
+
+---
+
+## Overview of files in repository
+
+```bash
+.
+├── CITATION.cff              # Citation metadata
+├── LICENSE                   # Open-source license
+├── README.md                 # Main project documentation
+├── assets/                   # Diagrams, plots, confusion matrices, LaTeX images
+├── cloud/                    # LambdaCloud configs and scripts
+├── data/                     # MathWriting, DeepMind MATH, CSAI, and related corpora
+├── docs/                     # Academic PDFs and source papers
+├── install_pixi.sh           # Environment setup script
+├── logs/                     # Training and evaluation logs
+├── models/                   # Saved model files (.h5, .keras, .json)
+├── modules/                  # DataLoader and utility code
+├── notebooks/                # All core modeling and evaluation notebooks
+├── pixi.lock                 # Pixi environment lock
+├── pixi.toml                 # Pixi environment definition
+└── sandbox/                  # Experimental code (dev branch)
+```
+
+---
+
+## Software Setup
+
+```bash
+# Preferred
+pixi install
+
+# Or use pip
+pip install -r requirements.txt
+```
+
+Dependencies:
+
+ * `tensorflow`==2.15, 2.18, 2.19
+ * `transformers`
+ * `gensim`
+ * `xgboost`
+ * `scikit-learn`, `matplotlib`, `numpy`, `pandas`
+ * `pixi`
+ * `python` >=3.12.11
+
+---
+
+## Data
+
+* Public data:
+  * [MathWriting Dataset](https://huggingface.co/datasets/deepcopy/MathWriting-human)
+  * DeepMind MATH (local generation)
+* Binary labels generated by shuffling MathWriting LaTeX IDs
+
+Run preprocessing via:
+
+```bash
+jupyter notebook notebooks/<notebooks>.ipynb
+```
+
+---
+
+## Training
+
+All model training is notebook-based. Start with:
+
+ * basemodel.ipynb for image classification
+ * model2.ipynb for text modeling
+ * model3.ipynb and model4.ipynb for classical approaches
+ * protopes.ipynb for GPT-based designs
+
+## Performance Evaluation
+
+Use:
+```bash
+jupyter notebook notebooks/comparemodels.ipynb
+```
+To reproduce visualizations, metrics, and ROC plots.
+
+---
+
+## Citations
+
+Tan & Le (2019). EfficientNet
+Vaswani et al. (2017). Attention Is All You Need
+Mikolov et al. (2013). Word2Vec
+Chen & Guestrin (2016). XGBoost
+Gervais et al. (2025). MathWriting Dataset
+Hendrycks et al. (2021). MATH Dataset
+Cocker, R. (2025). Version-Tab (Version 0.1.0) [Computer software]. https://github.com/rcghpge/version-tab
+
+---